Local-first · API · Optional MCP
A workstation that unifies desktop automation, vision & OCR, browser flows, and packaged skills—so teams can move from AI experiments to repeatable operations without betting everything on a single black-box chat.
Generative AI changed how people imagine work—but production still runs on scripts, APIs, desktops, compliance, and humans who need traceability and repeatability.
Models excel at language and ideas, yet they lack safe, consistent access to your screen, files, and legacy tools unless you deliberately engineer bridges. “Just ask the bot” rarely survives audit, outages, or second-shift handover.
Every new SaaS promises an API; every team still juggles spreadsheets, PDFs, green screens, and one-off Python snippets. Glue code multiplies. Knowledge lives in chat logs instead of versioned artifacts.
IT wants governance; operators want speed. Without a single place to prototype, record, package, and expose capabilities, “innovation” and “operations” pull in opposite directions.
Sensitive workflows often require on-machine execution, explicit user consent, and clear boundaries—not every step should traverse a third-party cloud.
Automation Skill Builder exists to shorten the path from “we need to automate this” to “we have a documented, invokable capability”—without forcing a single vendor narrative.
Plain language: we built a practical control room for automation and skill packaging—not another generic “AI platform” slide deck.
High-level capabilities of the application (exact feature set evolves with releases; see the in-app UI and manual for detail).
Design and run automations, manage favorites, browse marketplace-oriented skills, and configure the service from one place.
Screenshot, regional OCR, window management, and related building blocks for real-world UI automation.
Invoke capabilities programmatically; suitable for integration with your own schedulers, agents, or backends.
Workflows that help capture API-style interactions and translate intent into structured scripts where the product supports it.
Package automation artifacts for deployment contexts aligned with FastSkills / MCP-style consumption.
When enabled, expose selected OpenAPI operations as MCP tools so compatible assistants can call approved endpoints.
Summarized for operators and procurement. The user manual §2.1 is authoritative for version pins, file names, and CI details.
.paddlex).
Use a current browser for the local web UI (often http://127.0.0.1:8800).
No minimum Windows build is pinned in the repo; use 64-bit Windows 10 or later (or current Server) with a supported Python 3.10+ build.
Install requirements-windows.txt alongside the base requirements (includes pywin32 ≥ 306, PyAutoGUI, pynput, etc.). WSL counts as Linux for dependencies, not Windows.
No minimum macOS version is pinned; use a release where you can install Python 3.10+ and wheels from requirements-macos.txt (includes pyobjc and mac-specific packages).
Desktop automation needs Accessibility (and Input Monitoring if prompted) for your terminal, IDE, or packaged app—see the manual troubleshooting section.
Use requirements-debian.txt (same dependency tree as the main lockfile but without pyobjc). glibc-based 64-bit distros—especially Ubuntu / Debian—are the easiest path; others may need extra work for Paddle/PyTorch wheels.
Tray/splash GUI needs a graphical session (DISPLAY). For servers, run with --no-gui when you only need HTTP.
Packaged binaries are produced on GitHub Actions with windows-latest, macos-latest, and ubuntu-latest—see .github/workflows/build.yml.
Two lanes—so expectations stay clear whether you are evaluating, self-hosting, or running under a commercial agreement.
The software may be used and inspected according to its license and repository terms. Documentation, issues, and community discussion (where offered) are provided without warranty: no guaranteed response time, no SLA, and no obligation to implement feature requests.
Best for: tinkerers, internal pilots, and teams comfortable reading logs, updating dependencies, and owning their runtime.
For organizations that need onboarding, priority help, custom integrations, security review assistance, or deployment hardening, commercial support may be available from the maintainer or partners.
Scope, channels, and SLAs are defined in a separate agreement—not on this marketing page. Contact your vendor or project owner for entitlement and pricing.
This site is informational. It does not constitute a service contract. Always verify support terms in writing before relying on them for production.
For install, navigation, settings, scenarios, API/MCP notes, and troubleshooting, use the detailed user manual shipped with the repository.
requirements.txt plus platform files where provided). See System requirements for Windows / macOS / Linux notes.
python main.py); default HTTP is often
http://127.0.0.1:8800/ unless you override AUTOMATION_SERVICE_PORT.
Short answers; the manual remains authoritative for operational detail.
vscode-embedded-ui/ in the repo.